rdmsm4x — Tyrell resume + cross-project context filing

Session: 86816e7d-d320-481d-be49-47c476a4bd72 ("tyrell.app build kickoff", cwd ~/dev/tyrell) Span: 2026-08-26 18:49 → 19:05 EDT · rdmsm4x · Claude Opus 5 Budget posture at start and throughout: usage_status = hold (Anthropic 7-day vendor-reported 89→90%). All work done inline — no subagents, no fan-out, no escalation, per the delegation policy's hold rule.

Rich's instruction: "resume/complete this work — first document all contexts for other efforts/threads, unrelated to the primary goal here, and document/file/store them with their corresponding projects for reference/handoff that is automatically picked up and intelligently ingested … by the corresponding project/thread/task on its next run."

Part 1 — cross-project contexts filed (done first, as directed)

Used the existing ~/dev/todo intake system rather than inventing a parallel mechanism: one Markdown item per context with project: frontmatter, and the hourly LaunchAgent com.eastcoastscience.todoscan relays each into the owning project's ISSUES.md as an append-only [todo:<id>] line. That is the auto-pickup path Rich described — every project's agents already read ISSUES.md on start.

Four items written; all four verified relayed (three by the scanner firing on its own mid-write, the fourth by a manual scan_todo.zsh run):

Filed directly where no ISSUES.md exists for the scanner to target:

• SUPERSEDED-BY-TYRELL.md in concepts/llmstatescope, concepts/llm-work-reconciliation, and concepts/llm-wiki — each names what it contributed to Tyrell, where the work lives now, and how to route a gap (a ~/dev/todo item, not a restart here). Prevents a seventh partial rebuild of the same product.
• ~/dev/ISSUES.md — stray lowercase ~/dev/tyrell folder flagged (canonical repo is ~/dev/apps/Tyrell). Append-only write, because that file is under concurrent multi-agent write; a read-modify-write would have clobbered a peer.

Part 2 — Tyrell work completed

Found on resume: three files dirty from a prior session (Anthropic vendor-quota collector, an informational-bucket filter in the MCP, relayed todos). Verified: build green, 58 tests pass, collector runs clean, live vendor data correct. A concurrent session committed the collector work as b73854e mid-session.

Defect found and fixed — 4df3345: the informational-bucket rule had shipped in tyrell-mcp only. StatusFeed.pollUsageOnce graded any bucket carrying a used_pct, so an unrecognised Anthropic window (nimbus_quill, live today) reading high would light a red LLM BUDGET chip in the menu bar while usage_status, reading the same usage.json, said proceed — two verdicts from one document. The /usage detail panel was already correct (it lists every bucket with its caveat).

Root cause was duplication: four sites reduce the usage document, and the fix landed in one. Moved the rule to TyrellCore.UsageGrading (shared by both verdict surfaces) and pinned it with 5 new tests covering both failure directions — inventing headroom from a missing percentage, inventing exhaustion from an ungradable one. 63 tests green.

Sources/tyrell-mcp/MCPMain.swift was left uncommitted on purpose: a concurrent session is editing it and it carries their unfinished improvements. Behaviour is identical either way. Coordinated via fleet bus 20260826-190044-E1302CAC asking them to keep the UsageGrading call.

Shipped: universal signed bundle built, deploy_app_fleet.zsh installed + signature-verified on rdmbair15m5, rdmpw3275m, rdmbair13m5, jdmbair13m5. rdmpw3265m unreachable (already known offline; catches up next deploy). Local bundle synced and tyrellbar restarted — it had been running an 18-hour-old binary since 01:21, so without the restart the fix would have been "shipped" everywhere except the Mac Rich actually looks at.

Not done, deliberately

• No new work tranche started. Budget is hold; A5 iOS usage card, A6 watchdog usage alert, and Phase-1 items 2–5 remain queued.
• No Apple Note published via ~/scripts/notes_changelog.zsh. That script has an OPEN, unfixed data-loss hazard (~/dev/ISSUES.md, 18:50 today): AppleScript returns wrong counts rather than errors during the Notes cold-start stall, and line 205 deletes a folder it believes is empty. Used the Apple Notes MCP add_note instead — it cannot delete a folder.
• Nothing pushed, moved, renamed, or deleted — the 2026-08-24 fleet halt on structural change is still unreleased.

Rich's open decisions touched by this session

1. Rotate the dev.dataroo.net Basic-auth credential (filed to sites/dev.dataroo.net). The agent's refusal to use it does not undo the exposure.
2. ISSUES #21 still needs your call: quit ChatGPT.app on rdmpw3275m (may not stick — the proxy re-establishes from rdmsm4x) vs. scoping a ~/.codex runtime exclusion.
3. Fleet weekly Anthropic reads ~120% of the configured ceiling — the vendor endpoint now supersedes that estimate for grading, so this is less urgent than it was, but ~/.tyrell/usage-limits.json is still calibrated against one host.